November 2022 Summaries

In the Vespa Newsletter for November 2022, several key updates were highlighted, including performance improvements and new features. Multithreaded pre-filtering in Vespa 8.78.45 enhances approximate nearest neighbor queries by reducing latency through the use of the num-threads-per-search feature. The update also includes better hit estimates from parent document attributes, thanks to the use of fast-search with dictionary data in Vespa 8.84.14, which improves query plan efficiency. Additionally, new sample notebooks for training gradient boosting decision tree models using XGBoost and LightGBM were introduced, aimed at facilitating product search applications. Furthermore, Vespa Cloud, which had been available on AWS since 2019, is now also available in Google Cloud Platform (GCP) zones, expanding its accessibility for users.

In the continuation of a series on improving product search with learning to rank, this blog post explores the use of labeled relevance judgments to train deep neural ranking models, specifically focusing on neural cross-encoder and bi-encoder methods based on pre-trained language models. The dataset is split into training and development sets to avoid overfitting, and the models are evaluated using the Normalized Discounted Cumulative Gain (NDCG) metric. The cross-encoder model, which inputs both queries and documents into a Transformer model simultaneously, and the bi-encoder model, which encodes them independently, are compared. The results show that while both models improved over zero-shot baselines, the cross-encoder provided superior ranking performance. The blog also discusses the trade-offs between model accuracy, deployment costs, and the choice of product fields for model inputs. The post details the implementation of these models in Vespa for efficient product ranking and highlights the potential of combining neural models with lexical and statistical features for enhanced performance in future work.

Vespa Cloud, which has been operating in AWS zones since 2019, is now available in Google Cloud Platform (GCP) zones, specifically in the gcp-us-central1-f region, as announced by Kristian Aune, Head of Customer Success at Vespa.ai. This expansion allows users with existing workloads in GCP to integrate Vespa Cloud more seamlessly, potentially reducing data transfer costs, simplifying operations, and decreasing latencies by consolidating services within the same cloud provider. Users are encouraged to reach out if additional zones are needed, with a typical two-week ramp-up time for new zone availability.

In the blog post, the author explores the paradigm shift in product search ranking brought about by pre-trained language models like BERT, which have outperformed traditional statistical methods in data-rich environments. The focus is on using learning to rank techniques to improve product search performance, demonstrated through a large Amazon dataset containing complex search queries and relevance judgments. The post examines various zero-shot baseline ranking models, including traditional lexical models like BM25 and Vespa's nativeRank, semantic models using dense vector embeddings, and hybrid models combining lexical and semantic approaches. The findings reveal the challenges of applying dense vector models in zero-shot settings and highlight the effectiveness of lexical methods. The author suggests future posts will delve into training more sophisticated models using the labeled dataset, exploring advanced techniques like semantic similarity models, gradient boosting, and ensemble models within the Vespa platform.